What the EU AI Act requires, and how it affects SAP customers and Joule.
Analysis | ~12 min read | By J. Torre
In brief
- The EU AI Act regulates AI according to the risk of its use. Most AI faces no new obligations. A few uses are banned, some are classed as high risk and must meet detailed requirements, and some only need to be transparent.
- Duties depend on the role.
- The rules for high-risk systems apply from December 2027 and August 2028. The transparency rules and the supervision of general-purpose models already apply.
- SAP says it classifies its agents feature by feature. Agents in SuccessFactors that screen candidates, evaluate performance or allocate tasks are classed as high risk.
What the Act is and what it tries to solve
The European Commission describes the Act as the first comprehensive legal framework for AI. Its stated aim is to make AI in Europe trustworthy: protect safety and fundamental rights, and give companies clear rules so they can invest and sell across the single market.
In practice it addresses three problems.
1. Decisions that people cannot understand or challenge. Take a company that uses software to rank job applications. A candidate is rejected and asks why. If the vendor cannot explain how that ranking was reached, and the employer cannot show that it checked the result, nobody can tell whether the candidate was treated fairly. The same applies to a bank that uses software to set credit limits, or a school that uses it to grade exams. The Commission’s view is that existing law, such as data protection and anti-discrimination rules, gives some protection but does not require anyone to document, test and supervise these systems before they are used.
2. A few uses that the EU considers unacceptable in any case. Examples are scoring citizens on their social behaviour and using the score to decide who gets a service, or inferring employees’ emotions from a webcam. The Act bans these uses outright.
3. Different rules in each country. Without a common law, each of the 27 Member States could set its own requirements, and a company selling AI across the EU would have to meet all of them. A single regulation replaces that with one set of rules.
The Act does not regulate AI as a technology. It regulates uses, and the obligations grow with the risk that the use creates for people. The same software can be low risk in one use and high risk in another.
The four risk levels
To place a system in a level, ask what it is used for and what its output does. A tool that checks spelling has no effect on anyone’s rights. A tool that decides who gets a loan does. The Act turns that difference into four levels. The examples below are illustrative and based on the Act’s lists (they are not official).
| Level | What it means | Examples | What the Act requires |
|---|---|---|---|
| Unacceptable | Uses the EU considers incompatible with its values | A system that scores people on social behaviour and denies services based on the score; a tool that infers employees’ emotions from video at work | Banned |
| High | Uses that can significantly affect a person’s job, money, education, safety or legal position | Ranking and screening job candidates; assessing the creditworthiness of individuals; grading exams; allocating tasks and monitoring performance of employees | Allowed if the provider meets detailed requirements and the user follows rules of use |
| Transparency | Uses where people could be misled about talking to a machine or about what is real | A customer-service chatbot; an AI-generated image or video that looks real | The person must be told it is AI; generated content must be marked or labelled |
| Minimal | Everything else | A spam filter; demand forecasting; matching invoices to purchase orders; a spell checker | No new obligations |
The Commission states that most AI in use today is in the minimal group.
The use decides the level, not the technology. Consider one language model used in two ways. A finance team uses it to summarise supplier emails: minimal risk. An HR team uses the same model inside a tool that ranks candidates: high risk.
For high-risk systems, the provider must manage risks, use data of adequate quality, keep documentation and logs, give the user instructions, allow human oversight, and ensure accuracy, robustness and cybersecurity.
Models and systems are regulated separately.
A general-purpose AI (GPAI) model, such as an LLM (Large Language Model), can be used for many tasks: writing, summarising, programming. Its provider, that is, whoever develops it and places it on the market, has its own obligations: technical documentation, information for those who use it in their products, a copyright policy and a public summary of the training data. The most powerful models, with systemic risk, must also be evaluated, mitigate risks and report incidents.
An AI system is the specific product built on a model for a given purpose (use). The four risk levels apply to the system, according to what it is used for.
Therefore, the model provider meets the GPAI obligations, and the company that builds a recruiting tool on that model meets those of a high-risk system.
Who is responsible: providers and deployers
The Act assigns duties by role, and the same company can have different roles for different systems.
- The provider develops an AI system, or has it developed, and places it on the market or puts it into service under its own name.
- The deployer uses an AI system under its own authority in a professional context.
- Importers and distributors have smaller duties when they bring systems into the EU or pass them on.
In rough terms, the provider is responsible for the system being safe and documented, and the deployer is responsible for using it properly. For a high-risk system:
| Provider | Deployer | |
|---|---|---|
| Main question | Is the system designed, tested and documented to the required standard? | Is the system used as intended, with people supervising it? |
| Typical duties | Risk management; data quality; technical documentation; logging; accuracy, robustness and cybersecurity; instructions for use; conformity assessment; registration in the EU database | Use according to the instructions; assign trained staff to oversee; monitor operation; keep the logs under its control; inform workers and, in some cases, the people affected; assess the impact on fundamental rights where required |
Both must take measures to make sure their staff understand AI well enough to use it responsibly (AI literacy).
Examples of how the roles work
A model provider, a software vendor and a customer. A model provider develops an LLM and releases it. It has the GPAI duties. A software vendor, such as SAP, builds an assistant on top of that model and sells it as part of its product. The vendor is the provider of the assistant as an AI system and has the duties that match its risk level. A company that buys the product and uses it in its operations is the deployer.
An HR tool. A company uses an agent in its HR software that screens and ranks candidates. Recruitment is a high-risk area, so the agent is high risk. The vendor, as provider, must prepare the documentation, test the agent, enable logging and give instructions. The employer, as deployer, must use it as instructed, make sure a trained person reviews the rankings before decisions are taken, keep the logs and inform employees and their representatives where required.
A customer that changes role. A company takes a general-purpose agent meant for drafting emails and configures it, on its own initiative, to pre-select job candidates. This change of purpose turns the system into a high-risk one and can make the company the provider of that system, with the provider duties above. The same can happen if the company puts its own name on a system or modifies it substantially.
A finance team. A company uses an agent that matches invoices and proposes postings. It processes company data, does not decide anything about individuals’ access to services, and does not fall into a high-risk area. It is most likely minimal risk, with no specific obligations beyond AI literacy. If another agent assesses the creditworthiness of individual customers, that use is high risk and the analysis changes.
Why the role matters in practice. The role decides who owes what, so it should appear in contracts: what documentation and instructions the vendor provides, who keeps the logs, and who is responsible for human oversight.
Fines are proportionate to the breach. The maximum is EUR 35 million or 7% of worldwide annual turnover for prohibited practices, and EUR 15 million or 3% for most other obligations. For SMEs the lower of the two amounts applies.
Where things stand
| Date | What applies |
|---|---|
| 1 August 2024 | Entry into force |
| 2 February 2025 | Banned practices and AI literacy |
| 2 August 2025 | Obligations for general-purpose model providers |
| 2 August 2026 | Transparency rules; the Commission can enforce against general-purpose model providers |
| 2 December 2027 | High-risk systems in the listed areas (recruitment, credit and others) |
| 2 August 2028 | High-risk AI embedded in regulated products |
The original dates for high-risk systems were August 2026 and August 2027. A 2026 amending regulation, known as the Digital Omnibus on AI, moved them. The postponement reflects that the standards and guidance needed to comply were not ready. The same amendment also added a ninth banned practice (AI that generates non-consensual sexual content) and eased some obligations, such as AI literacy.
Since August 2026 the Commission’s AI Office supervises general-purpose models, and national authorities supervise other AI systems.
How AI model providers comply
The obligations for general-purpose model providers have applied since August 2025. The main route to compliance is the General-Purpose AI Code of Practice, published in July 2025. It is voluntary.
Signatories listed by the Commission include Amazon, Anthropic, Google, IBM, Microsoft, Mistral AI, OpenAI, Aleph Alpha, ServiceNow and Cohere. Meta declined to sign, citing legal uncertainty and obligations that go beyond the Act. xAI signed only the safety and security chapter. Not signing does not remove the obligations, but the provider must show compliance by other means. The list changes, so it should be checked on the Commission’s page before relying on it.
Providers must also publish a summary of the content used to train the model, using a mandatory Commission template. Models placed on the market before August 2025 have until August 2027.
For AI-generated content, the Commission published at the end of July 2026 a first list of more than 180 signatories of a voluntary code of practice on marking and labelling such content.
SAP and the AI Act
SAP’s role
SAP states that it can act as provider, deployer, purchaser or integrator of AI systems, and that it also develops its own foundation models. For customers, the most relevant case is the SAP Business AI portfolio, where SAP is the provider of features such as Joule and the customer is the deployer.
A Q&A published by heise/iX in July 2026 makes the same split and adds that the customer’s role can change. A customer that builds agents in Joule Studio and operates them under its own name, or makes them available to group companies, customers or third parties, may be regarded as a provider of those agents. Whether that leads to high-risk obligations depends on what the agent does.
Its main exposure appears to be as provider and deployer of AI systems that use third-party models, plus any models it develops itself. This should be confirmed on the Commission’s list.
What SAP says it does
SAP’s Responsible AI approach is organised around ethics, security and compliance:
- A global AI ethics policy based on the ten UNESCO principles, with an ethics office, an internal steering committee, an external advisory panel and a review committee for high-risk use cases.
- An ethics impact assessment for every AI use case, which classes it as minimal or no risk, high risk, or a red line. Red-line cases are stopped. High-risk cases need additional review and approval.
- ISO/IEC 42001 certification for AI management, which SAP says covers Joule, SAP AI Core and key Business AI products. This is SAP’s own statement. The certificate covers management processes, not the conformity of any single system with the Act.
- On the Act itself, SAP says it aligns its operations with the phased obligations and that a cross-functional team reporting to the Executive Board is building a classification process.
On the technical side, an SAP Community post from July 2025 describes how Joule relies on the customer’s existing role-based authorisations in the SAP applications. Customer data stays in the application tenant and is not stored in Joule, and third-party model providers do not store the data or train on it. SAP also offers the SAP AI Agent Hub, built on LeanIX, which gives an inventory of agents and models, risk classification and lifecycle governance. It supports a customer’s own inventory and classification work; it does not make the classification decision for the customer.
Joule Agents: SAP’s own position
In June 2026 SAP published a six-page customer briefing, Joule Agents: Governance, Auditability, and EU AI Act Compliance (version 1.0, 17 June 2026). It is the clearest statement we found of how SAP applies the Act to its agents. The points below are SAP’s position as written in that document; SAP states that it is not legal advice and that its views may change.
Agents are AI systems under the Act. SAP says Joule Agents meet the definition: they work with varying levels of autonomy and produce outputs that influence real environments, in this case ERP workflows and the business data they govern. Unlike rule-based automation, they reason over variable inputs. Unlike assistants, they pursue an objective across several steps without a separate human instruction for each action.
Classification is done feature by feature. A Joule Agent that performs or materially influences a function in a high-risk area is classed as high risk. In SAP SuccessFactors, agents support candidate screening and ranking, performance evaluation and task allocation, and SAP places these in the employment area. In finance and insurance, SAP says agents may influence decisions on access to essential services. The briefing reports that the confirmed high-risk classifications to date are concentrated in SuccessFactors, that the assessment of the portfolio is not finished, and that more classifications are expected. A feature can be exempt if it only performs a procedural task, improves the result of a human activity (for example grammar suggestions), or prepares information that does not determine an outcome. SAP says exemptions apply to individual features, must each be justified, and are reviewed as capabilities change. A customer’s use in a regulated sector can also bring an exempt feature back into scope.
Identity, permissions and audit trail.
- Each agent has its own identity, managed centrally by SAP, and its actions are logged in the same infrastructure as human user activity.
- When an agent acts for a user, its permissions are a subset of that user’s. It cannot exceed the user’s authorisation.
- The audit trail answers four questions: who acted, what they were authorised to do, what they did (permitted and blocked actions), and on whose behalf, including the delegation chain.
Human oversight. SAP says oversight is provided by the approval mechanisms already built into the ERP workflows. Users can review, override and reverse what an agent did, and a non-AI alternative for completing the task is always available. SAP’s AI ethics policy defines three oversight models (human-in-the-loop, human-on-the-loop and human-in-command) and the impact assessment decides which applies.
A limit SAP states itself. The technical standards that translate the Act into specifications are not expected before the end of 2026. Until they exist, conformity assessments cannot be finalised, and SAP says its current governance positions should be treated as interim. The briefing was published before the amending regulation entered into force and does not give dates; the later deadlines give more time, but the point about standards stands.
Two observations. First, SAP describes control through mechanisms that already exist (roles, approvals, logs). That is a reasonable basis for oversight and traceability, but whether it is enough for the high-risk requirements will be judged against standards that are not yet published. Second, the document is a vendor description and has not been assessed by an independent party.
Human resources is the area to look at first
Recruitment and worker management are high-risk areas in the Act. SAP’s own briefing confirms that SuccessFactors agents that support candidate screening, performance evaluation and task allocation are classed as high risk. For a customer using these features, SAP carries the provider obligations and the customer carries the deployer obligations: use according to the instructions, human oversight by trained staff, monitoring, log retention, and information for employees and their representatives. The briefing does not list which individual SuccessFactors features are classified or exempt, so customers should ask SAP for the classification of each feature they use.
What a CIO or CFO can do now
- List the AI in use. Include AI embedded in SAP applications, features that are switched on by default, agents built in-house and AI from other vendors.
- Establish the role for each case. Provider or deployer, and whether any in-house work changes the role.
- Classify the use. Check each case against the banned practices and the high-risk areas. HR, credit and access to essential services need attention first.
- Check transparency. Chatbots and generated content are subject to the transparency rules now.
- Ask vendors for evidence. For SAP: the classification of each feature you use, the instructions for use that deployers of high-risk systems must receive, and the status of conformity assessment.
- Plan for December 2027. The high-risk deadline moved, but the standards that define compliance are still being finalised.
Sources
- European Commission, AI Act
- Gibson Dunn, EU AI Act Omnibus agreement (May 2026)
- Help Net Security, EU begins enforcing AI Act (4 August 2026)
- heise/iX, Joule is running, compliance is not (27 July 2026)
- SAP, Responsible AI
- SAP, Joule Agents: Governance, Auditability, and EU AI Act Compliance (Customer Enablement, v1.0, 17 June 2026)
- SAP Community, Principles to Practice: Securing SAP Business AI
- Commission, General-Purpose AI Code of Practice
- WilmerHale, Mandatory template for AI training data summaries
